Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Break down full ICML reviewer responses into structured rebuttal units. Use when input contains reviewer summary/presentation/contribution/strength/weakness/question text and the goal is to split weaknesses/questions into granular R-style response items while preserving original wording for quoted issues.
.claude/skills/runtsang-stage1-icml-breakdown/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-06 | ✗→✓ | ▲ Improved | -30% | 0% |
| case-13 | ✓→✗ | ▼ Worse | -54% | 0% |
| case-22 | ✓→✗ | ▼ Worse | 118% | 0% |
| case-04 | ✓→✓ | ＝ Same ✓ | -71% | 0% |
| case-07 | ✓→✓ | ＝ Same ✓ | 0% | 0% |
Apply the shared template first:
skills/stage1/template/SKILL.mdThen apply only the ICML-specific overrides below.
Extract exactly these six keys (numbers only):
rating <- Overall Recommendationconfidence <- Confidencesoundness <- Soundnesspresentation <- Presentationsignificance <- Significanceoriginality <- Originalitysummary <- Summarystrength <- Limitations (keep compatibility with shared output key naming)Strengths And Weaknesses (extract only concern/criticism parts)Key Questions For Authors# Stage1 ICML Breakdown## Scores must contain exactly the six keys above.If "Strengths And Weaknesses" mixes praise and criticism in one bullet, keep praise in preserved context and split only actionable weakness concerns into atomic issues.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 12,076 | 7,260 | -40% | 1 | 1 | 0% | 2,048 | 1,673 | -18% | 0 | 0 | — |
case-02 | fail→fail | 7,253 | 9,810 | +35% | 1 | 1 | 0% | 1,288 | 1,989 | +54% | 0 | 0 | — |
case-03 | fail→fail | 11,552 | 13,529 | +17% | 1 | 1 | 0% | 1,993 | 2,892 | +45% | 0 | 0 | — |
case-04 | pass→pass | 17,653 | 2,782 | -84% | 1 | 1 | 0% | 2,688 | 768 | -71% | 0 | 0 | — |
case-05 | fail→fail | 7,956 | 2,215 | -72% | 1 | 1 | 0% | 1,160 | 650 | -44% | 0 | 0 | — |
case-06 | fail→pass | 10,651 | 5,640 | -47% | 1 | 1 | 0% | 1,642 | 1,149 | -30% | 0 | 0 | — |
case-12 | fail→fail | 10,694 | 1,831 | -83% | 1 | 1 | 0% | 1,783 | 552 | -69% | 0 | 0 | — |
case-07 | pass→pass | 3,685 | 2,023 | -45% | 1 | 1 | 0% | 632 | 631 | -0% | 0 | 0 | — |
case-08 | pass→pass | 9,568 | 3,239 | -66% | 1 | 1 | 0% | 1,892 | 950 | -50% | 0 | 0 | — |
case-09 | fail→fail | 6,377 | 5,729 | -10% | 1 | 1 | 0% | 1,011 | 1,202 | +19% | 0 | 0 | — |
case-10 | pass→pass | 13,225 | 4,660 | -65% | 1 | 1 | 0% | 2,393 | 1,091 | -54% | 0 | 0 | — |
case-11 | fail→fail | 4,027 | 1,983 | -51% | 1 | 1 | 0% | 641 | 529 | -17% | 0 | 0 | — |
case-13 | pass→fail | 11,274 | 2,913 | -74% | 1 | 1 | 0% | 1,662 | 757 | -54% | 0 | 0 | — |
case-14 | pass→pass | 15,167 | 1,919 | -87% | 1 | 1 | 0% | 2,510 | 551 | -78% | 0 | 0 | — |
case-15 | pass→pass | 5,289 | 1,937 | -63% | 1 | 1 | 0% | 779 | 597 | -23% | 0 | 0 | — |
case-16 | pass→pass | 12,497 | 1,662 | -87% | 1 | 1 | 0% | 2,510 | 546 | -78% | 0 | 0 | — |
case-17 | pass→pass | 9,620 | 2,250 | -77% | 1 | 1 | 0% | 1,399 | 551 | -61% | 0 | 0 | — |
case-18 | pass→pass | 7,683 | 1,761 | -77% | 1 | 1 | 0% | 1,082 | 578 | -47% | 0 | 0 | — |
case-19 | pass→pass | 17,582 | 7,541 | -57% | 1 | 1 | 0% | 2,680 | 1,446 | -46% | 0 | 0 | — |
case-20 | pass→pass | 7,186 | 2,785 | -61% | 1 | 1 | 0% | 1,186 | 779 | -34% | 0 | 0 | — |
case-21 | fail→fail | 6,199 | 3,054 | -51% | 1 | 1 | 0% | 896 | 857 | -4% | 0 | 0 | — |
case-22 | pass→fail | 16,565 | 26,974 | +63% | 1 | 1 | 0% | 2,511 | 5,463 | +118% | 0 | 0 | — |
case-23 | pass→pass | 11,767 | 15,734 | +34% | 1 | 1 | 0% | 2,048 | 3,176 | +55% | 0 | 0 | — |
case-24 | fail→fail | 7,796 | 14,513 | +86% | 1 | 1 | 0% | 1,432 | 2,859 | +100% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 24 cases were attempted. The headline lift of -4 percentage points is the difference between those two pass rates over the 24 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.